Cross-Domain Recommendation (CDR) aims to alleviate data sparsity by transferring user preferences from a source domain to a target domain. Existing methods typically build a shared preference bridge for all users, but often lack strong privacy protection, i.e., either exposing sensitive data or degrading performance via differential privacy. Moreover, they assume uniform transfer patterns, ignoring user-specific preference heterogeneity. To address these issues, we propose PFRCRP, a Parameter-Factorized Personalized Federated Recommendation framework for Cross-domain Rating Prediction. PFRCRP learns personalized bridge models under strict privacy guarantees by keeping all user data local through federated learning. We factorize user-specific bridges into common and personalized components via low-rank decomposition, and design a personalized aggregation mechanism to enhance collaboration among similar users. Experiments on three real-world datasets show that PFRCRP outperforms state-of-the-art methods in both recommendation accuracy and privacy preservation.